Files
stack/docs/scratchpads/gateway-install-ux-20260404.md
Jarvis ebf9517dd0
All checks were successful
ci/woodpecker/push/ci Pipeline was successful
ci/woodpecker/pr/ci Pipeline was successful
fix(mosaic): resumable gateway install + prominent admin token
The `mosaic gateway install` command had three UX problems:

1. Admin API token was silently saved to meta.json and never shown to
   the user — no way to retrieve it without reading the file.
2. If install crashed mid-way (e.g. daemon health-check failure), the
   next run prompted "Reinstall? [y/N]" and aborted on N, leaving a
   half-configured install with no resume path.
3. Health-check failures only pointed at `mosaic gateway logs` — user
   had to run a separate command to see the actual error.

Changes:

- **Prominent admin token banner** printed immediately after bootstrap
  creates the first admin user. Clear "save now, won't be shown again"
  warning.
- **Resumable install state machine.** Detects partial installs from
  meta.json + .env + mosaic.config.json + daemon state, and picks up
  where a prior attempt stopped instead of prompting. Fully set up
  installs now offer "re-run config wizard" with explicit warnings
  about what it does (regenerates .env, clears admin token, may change
  backend storage).
- **Inline log tail on health failure.** Last 30 non-empty lines of
  gateway.log printed automatically when the daemon fails to become
  healthy, so the user sees the underlying error without running a
  second command.
- **Corrupt-state detection.** If exactly one of .env / mosaic.config.json
  exists (from an interrupted prior install), refuses to guess and
  directs the user to `mosaic gateway uninstall`.
- **BETTER_AUTH_SECRET preservation** across config regeneration so
  existing Better Auth sessions aren't silently invalidated.
- **Admin token dropped on any config regeneration** (the wizard may
  point at a different backend; the old token is unverifiable).
- **Daemon stopped before config rewrite** so the live process never
  serves stale config.
- Bump `@mosaic/mosaic` 0.0.19 → 0.0.20.

Known follow-ups (noted in review):
- `--port 14242` as an explicit override cannot be distinguished from
  commander's default value; requires plumbing an `explicit` flag.
- No automated test coverage for the new state branches; requires
  mocking fs/readline/fetch/spawn.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-04 22:08:39 -05:00

48 lines
2.0 KiB
Markdown

# Gateway Install UX Fixes — 2026-04-04
## Context
User hit two publish-drift bugs on `mosaic gateway install` (fixed in #386 and
#389). On top of those, the install UX itself has three concrete problems:
1. Admin API token is generated by the bootstrap step and saved to
`~/.config/mosaic/gateway/meta.json`, but never shown to the user. There is
no way to retrieve it without reading the file or running `mosaic gateway
config` afterward.
2. When install crashes mid-way (e.g. daemon fails to become healthy), the
next run of `mosaic gateway install` prompts "Reinstall? [y/N]". Saying N
aborts, leaving the user with a half-configured install and no clear path
forward. There is no resume path.
3. Health-check failure only prints "Check logs: mosaic gateway logs" — forcing
the user to run another command to see the actual error.
## Goal
Make `mosaic gateway install` a single end-to-end, resumable command that shows
the admin token on creation and surfaces errors inline.
## Plan
- `install.ts`: detect partial state (meta exists, no admin token, daemon not
running) and skip already-completed phases instead of aborting.
- `install.ts`: add a prominent "Admin API Token" banner printed right after
bootstrap succeeds. Include copy-now warning.
- `install.ts`: on health-check timeout, read the tail of `LOG_FILE` and print
the last ~30 non-empty lines before returning.
- Keep `mosaic gateway config` as-is (view/edit env vars); this is not the
setup wizard.
- No new commands. No new flags yet.
## Out of scope
- Readline/stdin piping fragility (pre-existing, not related to user complaint).
- Refactor to a phased state machine (overkill for three targeted fixes).
## Verification
- Typecheck, lint, format (mandatory gates).
- Manual end-to-end: fresh install → confirm token displayed.
- Manual resume: delete daemon.pid mid-install → re-run → confirm it resumes.
- Manual failure: point ENV_FILE port to a used port → re-run → confirm log
tail is printed.